Definition of the question: Solar energy (a.k.a. photovoltaic energy) is one major source of renewable energy for future massive integration to the electricity grid. Electricity output by solar panels depends on solar irradiance reaching the surface (a.k.a. GHI, global horizontal irradiance). As a result, photovoltaic (PV) energy production varies with the fluctuation of solar irradiance. Accurate forecasting of PV energy production has been identified as one of most important issues for integrating PV into the electricity grid. Such forecasting is needed for different time horizons, from several minutes (nowcasting) to several days. Nowcasting heavily relies on extrapolation of current motions of clouds while the forecast for several days has to rely on operational numerical weather prediction (NWP) from weather agencies for changes of weather systems. However, as far as the forecast for 1-8 hours is concerned, such time horizon sits between the timescales suitable for nowcasting and for numerical weather forecast. As a result, there has been no consensus what approach is most suitable for such 1-8 hour forecast (a.k.a. intra-day forecast).
Research: The research team will take a hybrid approach for the intra-day forecast for PV production. The premise is that the data-driven algorithm is most promising and effective when it is guided by known physics and is then checked against not only the observations but also the simulations based on the most advanced physical models. The team will use available surface weather observations collected by the SJTU team, historical weather reanalysis archives and geostationary satellite imageries over the same period, as well as the WRF-LES model simulations to derive the algorithm. The WRF-LES model is one of the most advanced numerical models for simulation of cloud evolutions in atmosphere with a spatial resolution as fine as a few meters. Using simulations greatly increase the amount of data to be used for training the regression model. The research will take a two-step forecast approach. First, the weather regime will be classified based on large- scale weather conditions and satellite cloud imageries. Second, for each regime, both exogenous and endogenous variables will be used to train a regression model to forecast solar irradiance in 1-8 hours. The team will test two machine-learning models for the regression purpose: k-Nearest Neighbors (k-NN) model and Support Vector Regression Machines (SVM/SVR) models. Both models have been used before in solar forecast studies and we will study both and choose one with better performance. Together with surface temperature and wind speed forecast, such solar irradiance forecast will then be fed into an improved PV energy yield model to estimate the electricity output. The forecasts of both solar irradiance and PV output will be checked against the actual measurements that maintained by the SJTU team at the rooftop of Green Energy Laboratory building on the SJTU campus.